AutoML-Agent

This is the official implementation of AutoML-Agent: A Multi-Agent LLM Framework for Full-Pipeline AutoML (ICML 2025)

[Paper][Poster][Website]

Setup

Benchmark Datasets

Data Modality	Downstream Task	Dataset Name	# Features	# Train	# Valid	# Test	# Classes	Source	License	Evaluation Metric
Main Datasets
Image (Computer Vision)	Image Classification	Butterfly Image	224x224	4,549	1,299	651	75	Kaggle Dataset	CC0	Accuracy
		Shopee-IET	Varying	640	160	80	4	Kaggle Competition	Custom
Text (Natural Language Processing)	Text Classification	Ecommerce Text	N/A	35,296	10,084	5,044	4	Kaggle Dataset	CC BY 4.0	Accuracy
		Textual Entailment	N/A	3,925	982	4,908	3	Kaggle Dataset	N/A
Tabular (Classic Machine Learning)	Tabular Classification	Banana Quality	7	5,600	1,600	800	2	Kaggle Dataset	Apache 2.0	F1
		Software Defects	21	73,268	18,318	91,587	2	Kaggle Competition	N/A
	Tabular Clustering	Smoker Status	22	100,331	28,666	14,334	2	Kaggle Competition	N/A	RI
		Higher Education Students Performance	31	101	29	15	8	Research Dataset (UCI ML)	CC BY 4.0	RI
	Tabular Regression	Crab Age	8	53,316	13,329	66,646	N/A	Kaggle Competition	CC0	RMSLE
		Crop Price	8	1,540	440	220	N/A	Kaggle Dataset	MIT	RMSLE
Graph (Graph Learning)	Node Classification	Cora	1,433	2,708	2,708	2,708	7	Research Dataset (Planetoid)	CC BY 4.0	Accuracy
		Citeseer	3,703	3,327	3,327	3,327	6	Research Dataset (Planetoid)	N/A
Time Series (Time Series Analysis)	Time-Series Forecasting	Weather	21	36,887	10,539	5,270	N/A	Research Dataset (TSLib)	CC BY 4.0	RMSLE
		Electricity	321	18,412	5,260	2,632	N/A	Research Dataset (TSLib)	CC BY 4.0
Additional Datasets for SELA
Tabular (Classic Machine Learning)	Binary Classification	Smoker Status	22	85997	21500	143331	2	Kaggle Competition	N/A	F1
		Click Prediction Small	11	19174	4794	7990	2	OpenML
	Multi-Class Classification	MFeat Factors	216	960	240	400	10	OpenML
		Wine Quality White	11	2350	588	980	7	OpenML
	Regression	Colleges	44	3389	848	1413	N/A	OpenML		RMSE
		House Prices	80	700	176	292	N/A	Kaggle Competition

Usage

We recommend using conda environment.

conda create --name amla python=3.11
pip install -r requirements.txt

Run AutoML Development

1. Run the instruction-tuned LoRA adapter (Download Link) for Prompt Agent via vLLM. vllm==0.4.1 is strictly required to get correct parsed results.

Update: We have implemented an alternative OpenAI version via parse_openai(..)

HF_TOKEN="Your HuggingFace Token" CUDA_VISIBLE_DEVICES="0,1,2,3" python -m vllm.entrypoints.openai.api_server --model mistralai/Mixtral-8x7B-Instruct-v0.1 --enable-lora --lora-modules prompt-llama=./adapter/adapter-mixtral/ --tensor-parallel-size 4

2. Setup Prompt Agent and LLM backbone(s) in ./configs.py.

AVAILABLE_LLMs= {
"prompt-llm": {
"api_key": "empty",
"model": "prompt-llama",
"base_url": "http://localhost:8000/v1",
},
"gpt-4": {"api_key": "YOUR OPENAI KEY", "model": "gpt-4o"},
"gpt-3.5": {"api_key": "YOUR OPENAI KEY", "model": "gpt-3.5-turbo"},
}

3. Run chat with AutoML-Agent's Manager 🕴🏻!

fromagent_managerimportAgentManagerdata_path="agent_workspace/datasets/banana_quality.csv"# assuming the data is uploaded via web interface / APIuser_prompt="Build a model to classify banana quality as good or bad based on their numerical information about bananas of different quality (size, weight, sweetness, softness, harvest time, ripeness, and acidity). We have uploaded the entire dataset for you here in the banana_quality.csv file."manager=AgentManager(llm='gpt-4', interactive=False, data_path=data_path)
manager.initiate_chat(user_prompt)

Running in a Jupyter notebook is recommended. The generated output .py file will be in the agent_workspace.

Citation

@inproceedings{AutoML_Agent,
title={Auto{ML}-Agent: A Multi-Agent {LLM} Framework for Full-Pipeline Auto{ML}},
author={Trirat, Patara and Jeong, Wonyong and Hwang, Sung Ju},
booktitle={Forty-second International Conference on Machine Learning},
year={2025},
url={https://openreview.net/forum?id=p1UBWkOvZm}
}

License

This project is licensed under the CC BY-NC 4.0 license. Commercial use is prohibited.